Abstract
Since there is no individual approach that can be universally applied to effectively solve the hard problems of artificial intelligence and data analysis, hybrid systems are necessary to better tackle specific tasks by exploiting the advantages of different methodologies in a single framework. Based on known results of combining neural networks and rule-based systems, this work presents a hybrid system with the purpose of simplifying rule sets obtained from rule induction algorithms on classification problems without increasing the accuracy error. This is motivated by assuming that simplicity can lead to more understandable models and rule induction algorithms often provide an excessive number of rules necessary to classify future examples within a given accuracy error, even after pruning. Experimental evidence suggests effective gains on a benchmark of sixteen data sets. Experiments were also performed to detect the effect of different components of the proposed approach in achieving the results and so helping to explain why this hybrid system works.